Background Excessive proteins accumulation like collagen in the liver causes
Liver Fibrosis (LF)!. Biopsy is and has several limitations. Radiomics the gold standard approach to stage LF, but it is invasive is a method that extracts quantitative values (features) from medical images and improves diagnostic accuracy and can be considered as an alternative to biopsy?. The aim of this study is to determine the accurate stage of LF by developing three radiomics models based on extracted features from bi-parametric MR images (T۱-w images and ADC maps) by classifying the cases into one main class and two subclasses using
Machine Learning (ML) algorithms and assess their performance. Methods As LF is progressive, it was assumed that every liver segment has same LF stage?. Hence, three series of features were extracted from T۱-w images and ADC maps of determined ROI in top, hilum and bottom of the liver to triple the number of samples. The cases were split into the train and test sets with a ratio of ۷/۳. In order to balance the classes, SMOTE technique was used in train sett. After normalizing the quantitative features, Information Gain feature selection was used to select the best ten features from feature set. Support Vector Machine (SVM) algorithm with PUK kernel implemented to classify the cases firstly into insignificant LF and significant LF. Then according to result of the first model, second or third model was used to classify the cases into Normal and Primary LF or Advanced LF and cirrhosis. All three models analyzed the features both in single-parametric and bi-parametric forms. To evaluate the performance of the models, accuracy, sensitivity and specificity were calculated. ResultsThe ML results showed that in all three models, Bi-parametric form has higher performance in comparison with single-parametric forms. Accuracy of bi-parametric form in three models are ۰.۷۶, ۰.۸۵ and ۰.۸۱ respectively. The results are summarized in table۱.ConclusionThe primary results of our study demonstrate that bi-parametric MR features have a high potential in determining the stage of LF. In order to compensate for the low patient number, three series of features were extracted from images of each patient. Therefore, future studies should be performed with a large number of participants and assess the potential of radiomics analysis on other MR image types and multi-parametric forms.